1106

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 3, MARCH 2008

Signal Interpretation of Multifunction Radars:

Modeling and Statistical Signal Processing
With Stochastic Context Free Grammar
Alex Wang and Vikram Krishnamurthy, Fellow, IEEE

AbstractMultifunction radars (MFRs) are sophisticated

sensors with complex dynamical modes that are widely used in
surveillance and tracking. Because of their agility, a new solution
to the interpretation of radar signal is critical to aircraft survivability and successful mission completion. The MFRs three main
characteristics that make their signal interpretation challenging
are: i) MFRs behavior is mission dependent, that is, selection
of different radar tasks in similar tactic environment given different policies of operation; ii) MFRs control mechanism is
hierarchical and their top level commands often require symbolic
representation; and iii) MFRs are event driven and difference and
differential equations are often not adequate. Our approach to
overcome these challenges is to employ knowledge-based statistical
signal processing with syntactic domain knowledge representation: a signal-to-symbol transformer maps raw radar pulses into
abstract symbols, and a symbolic inference engine interprets the
syntactic structure of the symbols and estimates the state of the
MFR. In particular, we model MFRs as systems that speak
a language that can be characterized by a Markov modulated
stochastic context free grammar (SCFG). We demonstrate that
SCFG, modulated by a Markov chain, serves as an adequate
knowledge representation of MFRs dynamics. We then deal
with the statistical signal interpretation, the threat evaluation, of
the MFR signal. Two statistical estimation algorithms for MFR
signal are deriveda maximum likelihood sequence estimator to
estimate the system state, and a maximum likelihood parameter
estimator to infer the system parameter values. Based on the
interpreted radar signal, the interaction dynamics between the
MFR and the target is studied and the control of the aircrafts
maneuvering models is implemented.
Index TermsElectronic warfare, inside-outside algorithm,
GaltonWatson branching process, maximum-likelihood estimation, multifunction radar, stochastic context-free grammars,
syntactic modeling, syntactic pattern recognition.

I. INTRODUCTION

LECTRONIC support measure, a division of electronic

warfare, involves intercepting and interpreting radiated
electromagnetic energy for an operational commander to locate and identify radar sources, and evaluate their potential

Manuscript received November 26, 2006; revised July 16, 2007. This work
was supported in part by the Department of Defense Canada. The associate editor coordinating the review of this manuscript and approving it for publication
was Prof. Steven M. Kay.
The authors are with the Department of Electrical Engineering, University
of British Columbia, Vancouver, BC V6T 1Z4, Canada (e-mail: alexw@ece.
ubc.ca; vikramk@ece.ubc.ca).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSP.2007.908949

threats. The electronic support algorithm described in this

paper considers the self protection of the target from radar
threats, and a major component of which is the interpretation
of the intercepted radar pulses in terms of the possible radar
modes, such as search and track maintenance. In the current problem setup, because we focus on the target perspective,
the radar model is simplified by removing its multiple target
tracking capability, and we limit the scenario to having only
one multifunction radar in the proximity of the target.
In building electronic support systems to analyze radar signals, statistical pattern recognition has been used extensively.
Conventional radars could be characterized by fixed parameters
such as radio frequency, pulsewidth, and peak amplitude [1], [2].
For such radar characterizations, decision-theoretic approach as
in statistical pattern recognition is sufficient for solving signal
processing problems such as emitter identification and threat
evaluation. References [3] and [4] discuss template matching
of the intercepted radar signal against an EW library for both
emitter type and emitter mode identification. Histogram techniques are described in [5] to study temporal periodicities in
radar signals such as pulse repetition intervals.
However, modern radars, especially multifunction radars
(MFRs), makes the statistical pattern recognition approach inadequate. MFRs are radio-frequency sensors with beam-steering
antennas that are widely used in modern surveillance and
tracking systems, and they have the capability to perform a
multitude of different tasks simultaneously by multiplexing
them in time using short time slices [6]. The list of these
tasks includes search, acquisition, multiple target tracking, and
weapon guidance [7]. At the same time, they maintain low
probability of being detected and jammed.
The reasons for the inadequacy of the statistical pattern recognition are two folds. The first concerns with the exploding dimension of the feature space due to the versatility of the radar.
The possible variation of the radar parameters such as the carrier frequency and radar pulsewidth makes the statistical pattern
recognition infeasible. The second reason deals with the possible time varying feature space necessary for correct recognition. Because of the time multiplexing capability of the radar,
the underlying representation of the radar may need to vary in
order to capture the dynamics of the radar.
This paper considers a hybrid algorithm of both statistical and
syntactical pattern recognition techniques. The methodology is
to codify all a priori knowledge available and analyze observables within the context of the a priori knowledge. Because of
the success of formal language in codifying human language,
we propose to embody radar domain knowledge in a modified

1053-587X/$25.00 2008 IEEE

WANG AND KRISHNAMURTHY: SIGNAL INTERPRETATION OF MFRS

language representation, and implement signal interpretation as

a parsing operation through the radar pulses. In this representation, radar pulses are analogous to English letters, and control
rules of pulse generation to English grammar.
The origins of syntactic modeling can be traced to the classic
works of Noam Chomsky on formal languages and transformational grammars [8][11]. Among the many grammars and
languages that have been investigated for practical applications,
finite state grammar (FSG) and context free grammar (CFG),
as well as their stochastic counterparts, stochastic FSG and
stochastic CFG, are currently the most widely used classes of
grammars. The application of the grammars to syntactic pattern
recognition is covered in depth in [12]. In [13], stochastic
context free grammar (SCFG) is applied to study gesture recognition and monitoring of an online parking lot. In [14] and [15],
the dynamics of a bursty wireless communications channel is
modeled in SCFG. References [16] and [17] describe syntactic
modeling applied to bioinformatics and [18] and [19] apply
these models to the study of biological sequence analysis and
RNA. In addition, on a more related topic to our paper, SCFG
is studied in [20] and [21] as an alternative approach to plan
recognition.
In this paper, we model MFRs as Markov-modulated SCFGs
to take into account the MFRs mode dependent behavior, its hierarchical control, and the control law consisting of operational
rules. The more traditional approach such as hidden Markov and
state space models are suitable for target modeling [22], [23],
but not radar modeling. Traditionally, MFRs signal modes were
represented by volumes of parameterized data records known as
electronic intelligence (ELINT) [1]. The data records are annotated by lines of text explaining when, why and how a signal
may change from one mode to another. This makes radar mode
estimation and threat evaluation fairly difficult. In [24] and [25],
SCFG is introduced as a framework to model MFRs signal and
it is shown that MFRs dynamic behavior can be explicitly described using a finite set of rules corresponding to the production
rules of the SCFG. SCFG has several potential advantages that
follow.
i) SCFG is a compact formal representation that forms a
homogeneous basis for modeling and storing complex
system domain knowledge [12], [26], [27], and in which
it is simpler and more natural for the model designer to
express the control rules of MFR [24]. Specifying the production rules of the SCFG allows convenient modeling of
the human computer interface.
ii) SCFG is more efficient in modeling hidden branching processes when compared to a stochastic regular grammars
or hidden Markov models with the same number of parameters. The predictive power of a SCFG measured in
terms of entropy is greater than that of the stochastic regular grammar [28]. SCFG is equivalent to a multitype
Galton-Watson branching process with finite number of
rewrite rules, and its entropy calculation is discussed in
[29].
iii) The recursive embedding structure of MFRs control rules
is more naturally modeled in SCFG. As we will show
later, the Markovian type model has dependency that has
variable length, and the growing state space is difficult
to handle since the maximum range dependency must be
considered.

1107

In summary, the main results of the paper are as follows.

1) A careful detailed model of the dynamics of an MFR
using formal language production rules. By modeling the
MFR dynamics using a linguistic formalism such as a
SCFG, a MFR can be viewed as a discrete event system
that speaks some known, or partially known, formal
language [30]. Observations of radar emissions can be
viewed as strings from this language, corrupted by the
noise in the observation environment.
2) Novel use of Markov modulated SCFGs to model radar
emissions generated by MFR. The complex embedding
structure of the radar signal is captured by the linguistic
model, SCFG, and the MFRs internal state is modeled by
a Markov chain. This modeling approach enables the combination of the grammars syntactic modeling power with
the rich theory of Markov decision process.
3) Statistical signal processing of SCFGs. The threat evaluation problem is reduced to a state estimation problem.
Maximum likelihood estimator is derived based on a hybrid of the forward-backward and the inside-outside algorithm. (Inside-outside algorithm is an extension of HMMs
forward-backward algorithm [31].)
4) Parameterizing the MFR model with the targets maneuvering models, the interaction between the target and the
MFR is studied. The targets probing of the MFR in order
to find a maneuvering model that maximize its safety is
formulated as a discrete stochastic approximation problem,
and simulation study of the problem is performed.
The rest of the paper is organized as follows. Section II describes the multifunction radar in detail and its role in electronic warfare. Section III models the MFRs command generation mechanism, where the construction of the Markov chain
in terms of the MFRs goals and subgoals, and MFRs hierarchical control as a set of syntactic rules are detailed. Section IV
presents the threat estimation algorithm and the discrete stochastic approximation algorithm, and Section V provides the
numerical studies. Finally, Section VI concludes the paper.
II. ELECTRONIC SUPPORT AND MFR
Electronic warfare (EW) can be broadly defined as any military action with the objective of controlling the electromagnetic
spectrum [32]. An important aspect of EW is the radar-target interaction. In general, this interaction can be examined from two
entirely different viewpoints, that of the radar and of the target.
From the radars viewpoint, the goal is to detect and identify targets, and to maintain a firm track. From the targets viewpoint,
the goal is to protect itself from radar-equipped threat by interpreting intercepted radar emissions and evaluating their threat
(electronic support or ES). In this paper, the targets viewpoint
is the focus, and MFRs are the specific threat considered.
The approach taken in this paper to interpret the MFR signal
is knowledge-based. The raw radar signal is interpreted with
respect to a grammatical model that describes its characteristics; the characteristics of interest is the order of the events detected, and the event occurrence time is not of much importance.
The signal interpretation consists of two main components, a
signal-to-symbol transformer and a symbolic inference engine.
Fig. 1 illustrates the two components in the context of the ES
architecture, and a brief description of which is given here: The

1108

Fig. 1. Electronic support (ES) framework considered in this paper. The radar
signal emitted by the MFR is captured by the ES system on board the target
after being corrupted by the stochastic environment. The system consists of an
antenna, a signal-to-symbol transformer and a symbolic inference engine. The
signal-to-symbol transformer consists of a receiver/deinterleaver and a pulse
train analyzer, and its main purpose is to map raw radar signal to abstract symbols that are recognizable by the symbolic inference engine. The symbols are
identified as a and b in the figure.

receiver processes the radar pulses intercepted by the antenna,

and outputs a sequence of pulse descriptor words (PDW), where
a PDW is a data structure containing parameters such as carrier frequency, pulse amplitude, and pulsewidth of an individual
pulse. The PDWs are then processed by the deinterleaver, and
segregated according to their originating radar emitters. The
pulse train analyzer further processed the deinterleaved PDWs,
and classify them into abstract symbols called radar words. (See
Section II-A for definitions.) Finally, the symbolic inference engine analyzes the syntactic structure between the radar words,
interprets its threat level, and outputs the results on a pilot instrumentation panel.
Because the receiver, deinterleaver and pulse train analyzer
have been well studied, the signal-to-symbol transformer is not
covered in this paper, and we only focus on the symbolic inference engine. Using an analogy between the structural description of the radar signal and the syntax of a human language, a symbolic inference engine is said to contain the prior
domain-specific knowledge of the language MFRs speak.
The knowledge consists of the operational rules and constraints
captured by the radar analysts that are believed to be applied
in the generation of the radar signal for each specific mission
goal, and such knowledge allows the radar analysts to distinguish grammatical radar signal from ungrammatical one,
and to reason about the particular mission goal the MFR is executing. In todays modern radar systems, the operational rules
are often implemented with fuzzy logic or expert system [22],
and conventional mathematical formalisms such as differential
and difference equations are not effective in analyzing them. Instead, in order to compactly store the syntactic knowledge of
the MFRs language, formal language theory is applied, and the
MFR language would be fully specified by the establishment of
a grammar [27].
As far as ES is concerned, the optimal approach is to collect a corpus of radar samples, and induce the grammar directly
without human intervention. However, because of the degree
of complexity and potentially lack of data on the MFR signal,
grammatical induction approach is impractical. In this paper,
stochastic context free grammar is chosen to model the MFR
signal for each of its mission goal because of its generality over
the hidden Markov and state space models, and the existence

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 3, MARCH 2008

Fig. 2. Radar signal corresponds to different layers of radar command generation hierarchy. A radar task consists of a sequence of radar commands that
would best achieve a tactic goal, and each radar command can be mapped to a
certain catenation of radar words that MFR is to execute.

of algorithms for parameter estimation. The context-free backbone is constructed from the domain-specific knowledge of the
MFRs signal generation mechanism. Section II-A describes the
MFRs domain-specific knowledge that would be used to construct the model for knowledge-based signal processing.
A. MFR System Architecture and Its Signal Generation
Mechanism
Before discussing the MFR architecture, we begin by describing the radar signal that is generated by different layers of
the MFR command generation hierarchy. The list below begins
by the actual radar pulses generated by the MFR, to the software
objects that are scheduled by the MFR processor, and ends with
the radar policy that governs the scheduling process.
Radar word: A fixed arrangement of finite number of
pulses. For example pulses with a fixed pulse repetition
frequency.
Radar command: Catenation of finite number of radar
words that is optimized for extracting certain target information. Examples are target acquisition and nonadaptive
track.
Radar task: The three main radar tasks are search, target
identification and target tracking, and each is implemented
by a template of radar commands designed to achieve the
tactical goal.
Radar mode: The constraints or emphasis on the execution
of certain radar tasks due to the mission requirements or
resource allocations.
An example of the above radar signal is illustrated in Fig. 2.
The radar task and the radar commands in the example are selfexplanatory, and the letters and denote radar words. The
vertical bars represent radar pulses, and a particular arrangement
of them makes up the radar words.
Following the macro/micro architecture as described in [22,
Section 15.5.6 15.5.6], the generation of the radar signal is modeled by a MFR composed of four basic components:1 a situation
assessment, a radar manager, a command scheduler, and a radar
controller, which are illustrated in Fig. 3. The chain of commands starts with the situation assessment which provides evaluation of the tactic environment to the radar manager. The radar
manager evaluates the threat accordingly, and enters the appropriate radar task to the planning queue for scheduling. The radar
1The system architecture does not include multiple target tracking functionalities such as data association. The paper focuses on a single targets self protection and threat estimation, and thus models only the radar signal that a single
target can observe.

WANG AND KRISHNAMURTHY: SIGNAL INTERPRETATION OF MFRS

Fig. 3. MFR system architecture. The situation assessment provides the evaluation of tactical environment to the radar manager. The radar manager, based
on the evaluation, selects a radar task on which the command scheduler/radar
controller will operate. The command scheduler plans and preempts the tasks in
the planning queue depending on the radar load, and the moves the tasks fixed
for execution to the command queue. The radar controller maps the tasks in the
command queue to appropriate radar commands, which is retrieved by the radar
for final execution.

task consists of a sequence of macro radar commands, and the

commands can be repeated or preempted in the planning queue
by the command scheduler. The commands that are fixed for
execution are passed to the radar controller, where they will be
mapped to the appropriate radar words and retrieved by the radar
for execution.
In the rest of the section, we will discuss the operational details of each of the MFR components, and their relationship
to the macro/micro architecture. More specifically, the macro
sensor management, as described in [22], requires the MFR to
have three basic components: an operating scheme, a performance standard, and an adaptation procedure, and the micro
sensor management requires the MFR to be able to select combination of radar pulses that best accomplish the performance
requested by the macro tasks given the system status. We will
describe how each of the requirements are satisfied by the MFR
components.
The macro management is accomplished by the radar manager and the command scheduler. Radar manager sets the operating scheme and the performance standard for the MFR. It is a
finite state machine that transitions among a set of tasks, with the
transition probabilities determined by the radar mode. It sets the
guidance to which radar commands are to be created by mapping each radar task to a template of radar commands. The mapping can be mission dependent, and such dependency models
the performance standard. For example, a radar task Target
identification for an existing track, depending on the performance standard, may be mapped to an template of radar commands such as {Alert, Nonadaptive track, Range resolution 1}
or {Alert, Nonadaptive track, Range resolution 2}, where Range
resolution 1 and 2 differ in carrier frequency and the radar waveforms used.
The command scheduler models the adaptation procedure,
and the adaptation is modeled by the schedulers ability to plan
and preempt radar commands in the planning queue. The command scheduler processes the radar commands stored in the
planning queue sequentially, and it plans, if the current command requests it, by appending radar commands in the planning

1109

queue, and preempts by inserting commands in front of the current command. The planning and preempting will be discussed
according to some rules to be specified.
The micro sensor management, on the other hand, is accomplished by the radar controller. Similar to the command scheduler, the radar controller processes the radar commands in the
command queue sequentially and maps the radar commands to
radar words according to a set of control rules. Each radar command may be mapped to a multitude of different radar words
depending on the tactic environment, and the mapping will be
specified explicitly later in terms of the grammars productions
in Section III.
As a remark, the control is separated into the command
scheduler and the radar controller because of the MFR needs
to be both adaptive and fast [33]. The command scheduler
orders radar commands by time and priority, and stores them
in the planning queue for it allows real time rescheduling. On
the other hand, due to the systems finite response time, radar
commands in the planning queue are retrieved sequentially
and placed in the command queue where no further planning
or adaptation is allowed. The radar controller maps the radar
commands in the command queue to radar words and which
are retrieved by the radar for execution.
III. A SYNTACTIC REPRESENTATION OF MFR DOMAIN
KNOWLEDGE
In terms of natural language processing, we model the MFR
as a system that speaks according to a stochastic grammar,
and more specifically, we place the domain knowledge discussed in the previous section in a compact mathematical
formalism called the stochastic context free grammar. In
Section III-A, an overview of the formal language theory is
provided. In Section III-B, the radar manager, the command
scheduler and the radar controller are modeled, and the details
of the Markov modulated SCFG are provided. In Section III-C,
a well posedness issue of the grammatical model is discussed.
A. Formal Languages and Transformational Grammars
A formal language can be broadly defined as any set of strings
consisting of concatenations of symbols. The complete set of
distinguishable symbols in the language is known as the alphabet and is denoted here by . For example, an alphabet
, and one language over this alphabet might
might be
consist of all finite (or null) repetitions of the combinations
followed by either or ; in this language, the strings , ,
and
are valid strings but
is not.
The general notion of a formal language is impractically
broad. It is much more useful, and intuitive, to specify a
language in terms of its structural patterns. This is often accomplished by defining a grammar [8], [10], [11] sometimes known
in the literature as a transformational grammar. In grammatical
.
is a
terminology, a grammar is a four-tuple
finite set of nonterminal symbols, is a finite set of terminal
.
is a finite set of production
symbols, and
is the starting symbol. The grammars are
rules, and
divided into four different types according to the forms of their
production rules [8], [34]. Specifically, context free grammar
of the form
where
has production rules
and
; the superscript
indicates the set of

1110

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 3, MARCH 2008

all finite length strings of symbols in a finite set of symbols

, excluding the string of length 0. The rule
indicates
by . In addition, as
the replacement of the nonterminal
shown in [10], any context free grammar may be reduced to
Chomsky Normal form, and which has production rules of the
and
, where , ,
, and
form
. An example of context free grammar in the Chomsky
Normal form consists of the following elements:

where the bar separates the two production rules, meaning

may be mapped to either
or
that the nonterminal
. Starting from the nonterminal , the strings can be derived
by applying production rules to iteratively replace nonterminal
symbols with substrings. The preceding example admits the following derivations:

etc.
As a shorthand notation, the multiple derivation steps in the last
. Furtherderivation above may also be expressed as
more, please note that the notation is used to express production rules, and is used to represent derivation or replacement
of nonterminals in a string.
In addition, as is often the case, a certain amount of uncertainty exists in the process under study. In order to make the
model more robust, and also to capture the random effect in the
model, probabilities are added to the set of production rules .
Stochastic context free grammar is a four-tuple
with all elements identical to the context free grammar except
is a finite set of stochastic production rules. Let be a nonin
terminal in , the probability of its production rule
is denoted as
, and the probabilities must satisfy

where is the set of all right hand sides for in

. For example, the grammar given above may be converted into a stochastic one by assigning the following probabilities to the production rules

A Simple Example of MFR and Inadequacy of HMM: As

compared to conventional radars, MFRs are distinguished by
their ability to switch between radar tasks, and plan ahead their
courses of actions [33]. As an illustrative example showing the
correspondence between the grammar and the MFR, consider
and ii)
,
production rules of the form: i)
and
are nonterminals representing radar comwhere ,
mands in the planning queue and is a radar command in the
is interpreted as directing
command queue. The rule
the command scheduler to append to the command queue, and
in the planning queue. Similarly,
is interpreted as
delaying the execution of in the planning queue and insert
in front. Suppose the planning queue contains the radar command , a possible generation of the radar words is illustrated
in Fig. 4. (The figure also illustrates the mapping of the radar

Fig. 4. The figure illustrates a possible realization of the scheduling process

represented by a grammatical derivation process. B and C are nonterminals
and b is a terminal. The triangle represents the mapping of the radar command
b to the radar words, y and w, by the radar controller.

commands to some radar words by the radar controller.) It can

be seen that as long as the command queue commands appear
only to the left of planning queue commands in the rule, the
command queue and the planning queue are well represented.
In addition to the interpretation of the production rules, another important property is their generative power, and why a
more established method such as hidden Markov model cannot
be used. As shown in [35], the rules of the form i have the
syntax of regular grammar and they can be used to represent
hidden Markov models, i.e., stochastic regular grammar. The
rules of the form ii, on the other hand, have the syntax of context free grammar. In other words, the MFR grammar has rules
that strictly contain regular grammar (rules of the form ii cannot
be reduced to i), and, thus, the MFR grammar cannot be sufficiently modeled by HMM. The production rules presented in
this example is a self-embedding context free grammar and it
cannot be represented by a Markov chain [10]. A context-free
grammar is self-embedding if there exists a nonterminal such
that
with
. For the rules presented,
self-embedding property can be shown by a simple derivation

In addition to the self-embedding property, HMM is not suitable because the radar controller may generate variable length
radar words. If HMM is to model the radar words, the Markovian dependency may be of variable length. In this case, maximum length dependency needs to be used to define the state
space, and the exponential growing state space might be an
issue. Furthermore, for sources with hidden branching processes
(MFRs), stochastic context free grammar is shown to be more
efficient than HMM in the sense that the estimated SCFG has
lower entropies [28].
B. A Syntactic Model for a MFR Called Mercury
In this subsection, because the MFR domain knowledge is
application dependent, for illustrative purpose, the grammatical
representation is discussed based on a particular type of MFR
called Mercury (The declassified version of the Mercurys textual intelligence report can be found in [36]). The output of the
MFR is modeled by a set of terminals, and the hierarchical command generation mechanism is modeled by a set of production
rules that map the top level radar tasks to radar commands, and
from radar commands to radar words.

WANG AND KRISHNAMURTHY: SIGNAL INTERPRETATION OF MFRS

TABLE I
LIST OF MERCURY RADAR COMMANDS AND THEIR CORRESPONDING RADAR
WORDS

The MFR grammar is

.
is the set of radar tasks.
and
are identical sets of
radar commands available to the MFR, and they are differare the commands
entiated only by their residing queues;
are in the command queue.
in the planning queue and
is the set of production rules mapping
to
.
is the set of production rules mapping
to
, where
is the set of radar words. In SCFG,
is the starting
is a Markov chain
symbol, however, in our formulation,
. The output of the Markov
with state space defined by
and it is the starting symbols for . Specific
chain is in
consists of nine distinct
to Mercury, the set of radar words
. The set of available radar commands is
elements
{Three-word search, Four-word search, Acquisition, Nonadaptive track, three stages of Range resolution, Track maintenance,
Fine track maintenance}, and it is written in shorthand as
,
or c denoting
or
respectively. Table I lists
where
the radar commands and their corresponding radar words.
The Mercurys grammar will be introduced according to the
framework depicted in Fig. 3. The radar manager is modeled as
, the command scheda Markov chain whose state space is
(self-embedding),
uler is represented by the production rule
and the radar controller, introduced along with the effects of the
stochastic channel, is modeled by the production rule . We
will describe each MFR component in detail.
1) Radar Manager: The radar manager, for each time
period, determines the overall task or tactical goal the MFR
is to accomplish. The time evolution of the radar manager
is modeled as a Markov chain, and its state space,
, is defined based
denote
on the major radar task categories [22]. Let
, is a three state
discrete time. The state of the MFR,
discrete time Markov chain. The output of each state is defined
by templates of radar commands that specify the type and the
order of the radar commands the MFR is to complete in order

1111

TABLE II
LIST OF TARGETS MOTION MODELS

to accomplish the tactical goal. The templates for the states are
expressed in the production rules listed here.
;
Search for new targets
Target identification for existing tracks
;
.
Track update for existing tracks
Each state may output multiple templates and they are separated
by bars. Different templates are characterized by their computational cost and accuracy, and their selection is modeled probabilistically.
,
Define the transition probability matrix as
, and and are MFR
where
states in . The transition of the MFR is assumed to be driven
by the interaction between the MFR and targets. For example,
if the target is far away from the MFR and flies with constant
velocity, the probability of the MFR jumping to Track update
for existing tracks might be low. On the other hand, when the
target is close and shows high maneuverability, the probability
of being tracked might be higher because MFR would allocate
more resources to it.
In order to characterize the interaction between the MFR and
a target, the target behavior pattern is described first. A target
, where refers to its kinematics
state process is
and is a staircase-type trajectory indicating its motion models
such as constant velocity model [37]. In this paper,
denotes distance of the target with respect to the MFR, and
is an indicator vector featuring the motion model in
which the target is maneuvering. The dependency between the
MFR and targets is established by parameterizing the transition
.
matrix with
Table II lists the values of and their corresponding motion
models. The list of representative motion models are used in
[38] to study the benchmark tracking problem. The first model,
constant velocity model, characterizes the periods of nonmaneuverability, and it is described in [39]. The other two models are
to account for target maneuvers. The time correlated acceleration model is first proposed in [40] and the horizontal turn model
is described in [41].
Because of its generality and utility interpretation, Logit
model is selected to parameterize the transition matrix. Let
be the probability of the MFR system to move up
is the probability of the MFR system
(down) a state and
remaining in the current state. The probabilities are illustrated
in Fig. 5 and they are shown as follows:

1112

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 3, MARCH 2008

TABLE IV
PRODUCTION RULES OF MERCURYS RADAR CONTROLLER

Fig. 5. MFR states and transition probabilities.

TABLE III
PRODUCTION RULES OF MERCURYS COMMAND SCHEDULER

where , and are vectors of regressor parameters. The justification of the logit model is given in Appendix A.
2) Command Scheduler: The command scheduler models
the MFRs ability to plan and to preempt radar commands based
on the radar task and the dynamic tactic environment. With the
template of radar commands in place, the main operation of
the command scheduler is to implement the scheduling of radar
commands in the command queue and/or the rescheduling of
commands in the planning queue. The operational rules for the
scheduling and rescheduling could be constructed based on a
and
small set of basic rules. Suppose
, the basic control rules that are available to the command scheduler are listed.
Markov and
Adaptive and
Terminating and
The interpretation of the rules follows the example given at
the end of the previous subsection. A rule is Markov if it sent a
radar command to the command queue, and re-scheduled either
a same or a different radar command in the planning queue. A
rule is Adaptive if it either preempted a radar command for another radar command or if it scheduled a radar command ahead
of time in the radars time line after the current command. A
rule is Terminating if it sent a radar command to the command
queue without scheduling any new commands.
The significance of the Markov rule is obvious. It represents
the completion of one radar command and the scheduling of
another. The two adaptive rules model the MFRs ability to: i)
Preempt and ii) Plan the radar commands. The preempt rule is
, where the command is preempted when a higher
priority task enters the queue. On the other hand, the plan rule
, where the command is scheduled ahead of time.
is
The terminating rule reflects the fact that the queues have finite
length, and the grammatical derivation process must terminate
and yield a terminal string of finite length. Applying the basic
could be
control rules to the templates, the production rule
constructed. With some constraints in place, the complete set of
rules is listed in Table III.
3) Radar Controller and the Stochastic Channel: The radar
command is mapped to the radar words by the radar controller,

and the words could be corrupted by the stochastic channel before its intercepted. Here, production rules of the radar controller are devised, and the effect of the stochastic channel is
incorporated.
The production rules of the radar controller are derived from
visual inspection of the radar commands listed in Table I. The
syntactic structure of the radar commands are captured by
defining the nonterminals and their corresponding production
rules. We begin by defining the triplets as follows:

and blocks of four words

Furthermore, we introduce two new nonterminals

The nonterminals introduced specifies the complete set of the

production rules for the radar controller.
Based on the radar controllers production rules, the effects
of the stochastic channel could be easily incorporated. For each
and the production
radar word , define a new nonterminal
rule

where
is a
would be corvector of probabilities indicating how likely
rupted and intercepted as one of the other radar words. When
compiled together, the complete set of production rules are
specified and they are listed in Table IV. As will be illustrated
in later sections, the probabilities of the production rules could
be estimated based on training data. In addition, since each
is a pulse train, a pulse train analysis can be conducted to assign
[42].
prior probabilities to the channel probabilities

WANG AND KRISHNAMURTHY: SIGNAL INTERPRETATION OF MFRS

1113

Fig. 6. A string of radar words are intercepted by the MFR, and the signal interpretation problem is, based on the domain specific knowledge on the MFRs control
hierarchy, how to infer the tasks MFR is performing from the radar words. Task 1 is searching for new targets, task 2 is target identification for existing tracks, and
task 3 is track maintenance for existing tracks.

C. Well Posedness of the Model

One practical issue of modeling with SCFG is that the signal
generated by radar systems has finite length, and this finiteness
constraint must be satisfied if the model is to be stable. In addition, the finiteness criteria provides a constraint on the SCFG
model parameters, which may be used as a bound on the parameter values. We discuss this point by first defining the stochastic
mean matrix.
, the stochastic mean matrix
Definition: Let
is a
square matrix with its
th entry being the
expected number of variables resulting from rewriting
Fig. 7. Inside and outside probabilities in SCFG.

where
is the probability of applying the production
, and
is the number of instances of in
rule
[43].
The finiteness constraint is satisfied if the grammar in each
state satisfies the following theorem.
is less than one, the
Theorem: If the spectral radius of
generation process of the stochastic context free grammar will
terminate, and the derived sentence is finite.
Proof: The proof can be found in [43].
IV. STATISTICAL SIGNAL INTERPRETATION OF THE MFR
SIGNAL AND CONTROL
Given the MFR knowledge representation as discussed previously, we are now in the position to describe the symbolic inference engine. (Recall the ES framework in Fig. 1.) The input to
the engine is a batch of noisy radar words stored in a track file,
and the aim is to extract the embedded syntactic pattern that is
described by the domain specific knowledge. Fig. 6 illustrates
the inference problem we are to solve. In general, with such an
assumption, any pattern recognition technique is automatically a
signal interpretation technique. Specific to our case, because the
knowledge is stored as a Markov modulated SCFG, a hybrid of
the inside-outside and the forward-backward algorithm will be

used. In this section, we describe the state estimation algorithm

with the assumption of complete system knowledge (known parameter values) in Section IV-A, and the application of EM algorithm to estimate the system parameters in Section IV-B. In
Section IV-C, we extend the estimation algorithm to the control
of the targets maneuvering models.
Notation: The following notation will be used throughout the
be the (unknown) state
section. Let
(See Section III-B-1), and
sequence, where
be the intercepted radar commands. Each
is a string of concatenated terminal symbols
is the length of . It is convenient to
(radar words), and
introduce the following variables:
;
forward variable:
backward variable:
;
;
Inside variable:
outside variable:
where
is the subsequence of terminals from
position of
to
position, and
is the nonterminal
that de, or
. Fig. 7 illustrates the inside and outside
rives
probabilities. (Details of forward and backward algorithms can
be found in [44], and inside and outside in [28].)

1114

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 3, MARCH 2008

A. Maximum Likelihood Estimation of MFRs State via Viterbi

and Inside Algorithms
The

estimator

of

MFRs

state at time
is
, and which could be
computed using the Viterbi algorithm. Define
, the
Viterbi algorithm computes the best state sequence inductively
as follows:
, for
.
1) Initialization:
2) Induction:

where

the

complete-data

likelihood

is
.
In order to facilitate the discussion of the EM algorithm, the
following two variables are introduced

and

The Expectation step of the EM algorithm yields the following

equation:

3) Termination:
4) Path backtracking:

.
, for

where
is the output probability of the string generated
by the grammar . An efficient way to calculate the probability
is by the inside algorithm, a dynamic programming algorithm
that inductively calculates the probability.
, inducThe inside algorithm computes the probability,
tively as follows:
.
1) Initialization:
2) Induction:

for
.
.
3) Termination:
Running both the Viterbi and the inside algorithms, the posteriori distribution of the states given the observation could be
computed.

where
can be computed using inside
and outside variables [35]. The Maximization step of the EM
algorithm could be computed by applying Lagrange Multiplier.
Since the parameters we wish to optimize are independently
separated into three terms in the sum, the three terms are the estimates of the prior distribution, the transition matrix, and the
production rule probabilities, we can optimize the parameter
term by term. The estimates of the probabilities of the production rules can be derived using the first term of the equation, and
the updating equation is

B. Model Parameter Estimation Using EM Algorithm

In Section IV-A, MFRs state estimation problem was discussed assuming complete knowledge of the system parameters,
i.e., the Markov chains transition matrix and the SCFGs production rules. In reality, such parameters are often unknown. In
this subsection, EM algorithm is applied for parameter estimation and it is discussed in detail in [45].
be the incomplete data, and let
be the
Let
states,
missing (or hidden) data. For a Markov chain with
and
rule

is the number of counts the production

is applied in deriving
with grammar . Let
be the model pais the set of production rules probrameters, where
abilities for grammar . The EM algorithm iteratively computes
the maximum likelihood parameter estimates by computing

Similarly, the updating equation of the transition matrix

is

Under the conditions in [46], iterative computations of the expectation and maximization steps above will produce a sequence
of parameter estimates with monotonically nondecreasing likelihood.
C. Optimization of Target-MFR Interaction Dynamics
Based on the interpretation of the radar signal and the interaction dynamics between the MFR and the target, autonomous
control of the aircrafts maneuvering model is devised in this
subsection. Recall the Target-MFR interaction as discussed in

WANG AND KRISHNAMURTHY: SIGNAL INTERPRETATION OF MFRS

Fig. 8. The selection of maneuvering model induces a particular radar mode.

The mode is observed indirectly from the intercepted radar pulses and its threat
evaluated. Based on the evaluation, the control strategy selects maneuvering
models such that the ownship safety is maximized.

Section III, where each maneuvering model triggers a particular radar mode, and the mode is characterized by the transition probabilities of the radar tasks. With this assumption, the
maneuvering model selection is formulated as an optimization
problem of finding an efficient adaptive search (sampling) plan
with the objective of staying in the safest mode most often,
and the problem setup is illustrated in Fig. 8.
indexes the sequence of
Let the discrete time
be the
maneuvering models selected by the aircraft. Let
single performance measure, the MFRs average occupancy in
track mode when the target is maneuvering in model , and
which can be computed from the stationary distribution of the
estimated Markov chain. The aim is to find such that

where
is the set of all possible maneuvering models. The
model selection is not straightforward because the performance
of the maneuvering cannot be evaluated analytically, and it
must be estimated or measured based on the intercepted radar
pulses. We treat this problem as a discrete stochastic approximation problem. The problem is also called the multiarmed
bandit where the aim is to find the best slot machine out of
a finite number of such machines. Other approaches such
as multiple comparison also exist [47], but this approach is
preferred because of its ability to adapt to slowly time-varying
radar conditions.
Two discrete stochastic approximation algorithms will be applied, and their detailed description can be found in [48]. The
target begins in an arbitrarily chosen motion model, and probabilistically explore the model space. The idea is to implement
an efficient adaptive sampling plan that allows one to find the
maximizer with as few samples as possible by not making unnecessary observations at nonpromising models. The following
is a sequence
notations are used in the algorithms.
of maneuvering models generated by the algorithm that can be
thought as the state of the algorithm at time 1. It is convenient
to a sequence of unit vectors
where it has 1
to map
, and zeros elsewhere. In adin the th component if
denotes
dition, let
the empirical state occupation probability measure, where
gives the number of elements in the set and
is a counter
that measures the number of times the state sequence visits the
is the estimate of the optimal mode genstate . Finally,
erated by the algorithm at time . It is the main output of the

1115

algorithm and it is used to control the aircrafts mode changes.

The two algorithms are summarized here.
Aggressive Search:
, select initial state
.
1) Initialization: At time
,
for all
,
. Set
Set
.
, compute
2) Sampling and Evaluation: Given the state
. Generate a candidate state
from
according to a uniformly distributed random variable.
.
Compute
, then set
3) Acceptance: If
; otherwise set
.
4) Adaptive filter for updating state occupation probabilities:
Update state occupation probabilities

with the decreasing step size

, where
cator function.
5) Update estimate of optimal radar mode: If

is indi-

then set

; otherwise, set
. Set
and go to Step 1.
Conservative Search:
, initialize state
-di1) Initialization: At frame time
to zero, and
(vector
mensional vectors
.
of ones). Select initial state
, generate,
2) Sampling and Evaluation: Given the state
, and
as in Step 1 of Aggressive Search,
. Update the accumulated cost, occupation times
and average cost as

3) Acceptance: If
, set
; otherwise set
.
. Set
4) Update estimate of optimal radar mode:
and go to Step 1.
The aggressive search explores the model space by jumping
between the models as a irreducible Markov chain, and it does
almost
not converge. However, it is shown in [48] that
surely, meaning the algorithm spends most time at the global
maximizer than any other state, and it is consistent. On the other
hand, the conservative search converges almost surely to the
globally optimal model. The convergence analysis of the conservative search holds for any size of the maneuvering model
sequence, as long as its greater than 0, where the aggressive
search requires long sequence. In addition, one advantage of

1116

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 3, MARCH 2008

the aggressive search is that, if we keep the step size constant

for both algorithms to make them adaptive to time-varying parameters, it is faster than the conservative search because it aggressively explore the state space. The numerical studies of the
algorithms are discussed in the next section.

TABLE V
THE SOURCE AND ESTIMATED PARAMETER VALUES OF THE MARKOV
MODULATED SCFG

V. NUMERICAL STUDIES OF THE ALGORITHMS

A software testbed is implemented in C++ for MFR signal
simulation and interpretation. In this section, the data structure
used to implement the algorithms, and some numerical results
will be discussed.
A. Implementation of the Software
The grammatical derivation process requires recursive embedding of terminals, repeated readings of nonterminals and
modification of the output string. In order to have efficient repeated memory access, the production rules and their probabilities are both stored as a map data structure indexed by nonterminals, and with their right hand sides implemented with linked
lists. In addition, the nonterminals and the terminals are stored
as vectors, and the starting symbol as a string. With this setup,
the grammatical derivation can be easily implemented by repeatedly accessing and joining the linked lists of the production
rules. In addition, because any context free grammars can be reduced to Chomsky Normal Form [10], the testbed is written to
accept only grammars in Chomsky normal form.
B. Model Complexity and Its Modeling Power
Here we describe briefly several implementation issues of
our testbed and the possible remedies. The major implementation issue of the testbed is with the inside-outside algorithm:
the computation complexity of the algorithm and the number of
local maxima in the likelihood function. Suppose the MMSCFG
has states, and the states are represented by a grammar with
nonterminals. Suppose further that the observation sequence has
radar words
length , and each observation has, on average,
. The average case complexity of each iterafor
tion of the EM parameter estimation algorithm is
(The complexity of the inside-outside algorithm for radar words
[35]), where
. However,
of length is
because the inside and outside algorithms could be run against
the data independently, parallel computation is possible and the
computation time could be reduced substantially. In order to
deal with the local maxima problem, one of the approaches is to
pick the initial parameter value more cleverly with pretraining
method introduced in [28], where significant computational savings is recorded and EM typically converges to the global maximum.
One important implementation detail regarding the modeling
power of the SCFG is its predictive power against branching
processes. In [28], study is done to compare the SCFG and the
HMM on their capability in modeling branching processes in
terms of entropy argument. In their study, a SCFG and a HMM
model are inferred against simulation data from a branching
process, and it is observed that the estimated SCFG consistently
has lower entropy than the estimated HMM model. Since our
MFR grammar is a multitype Galton Watson branching process,
SCFG has higher predictive power than HMM.

C. Numerical Results of the State and Parameter Estimation

In this subsection, the state and the parameter estimation
algorithms derived in Section IV-A and -B are evaluated against
simulation data. The model parameters such as the transition
probabilities and the production rule probabilities are estimated
and, based on the estimated values, the hidden state sequence
is inferred. For simplicity, the MFR is characterized by a
subset of the MFR grammar developed. The set of nonter, and the set of terminals is
minals is
. The grammars used in the numerical
studies are shown in Table V in its Chomsky Normal form,
and they characterize two different range resolution algorithms
with different performance standards. Because the grammar
is reduced, only two Markov states are considered, and the
templates used to define the states are identical except their
production rule probabilities. The Markov transition matrix is
assumed fixed in this study. Fig. 9 shows the evolution of the
likelihood values from the parameter estimation algorithm, and
the state estimation error probability with the parameter values

WANG AND KRISHNAMURTHY: SIGNAL INTERPRETATION OF MFRS

Fig. 9. The left figure shows the likelihood values obtained from iterating the
parameter estimation algorithm, and the right figure is the state estimation error
probability with the parameter values for each iteration of the algorithm.

for each iteration of the algorithm. The final estimated parameter values are listed in Table V, and it can be seen that the
estimated parameter values are very close to their true values.
In addition, the effect of the initial values on the parameter
and state estimation is also studied. We initialize the estimation
algorithms with values of different square-distance from the true
values, and run the parameter and state estimation algorithms.
It is found that the algorithm is not sensitive to the initial values
of the transition matrix, but it is sensitive to the initial values
of the production rule probabilities. One observation is that if
the grammars of different states are initialized too close to each
other, the Markov chain degenerates into an i.i.d. sequence and
the estimation algorithm updates only one state instead of two.
For transition matrix along, the rms (root mean squared) error of
the initial values to the true values, and of the estimated parameter values to the true model parameters are listed here. The rms
error of the estimated model parameters are very close to each
other despite of the differences in the initial values. Moreover,
the state estimation error probabilities of the cases shown in the
table at the bottom of the page all approach zero.
D. Numerical Results of the Autonomous Selection of
Maneuvering Models
In the second numerical study, we look at the interaction between the radar and the target maneuvers, and how the target
selects its maneuvering models according to discrete stochastic
approximation algorithms introduced in Section IV-C. The scenario is illustrated in Fig. 10. We assume that the target intends
to follow a circular path, circumventing the MFR, to reach a location labeled by X in the figure. The path is planned before
the mission, and the target switches between its maneuvering
models to maximize its safety.
In this paper, the target is assumed to be able to maneuver
in four different motion models, and the MFR would respond
with four corresponding radar modes characterized by their
Markov modulated SCFG representations. Because the targets

1117

Fig. 10. The scenario of the numerical study sets a target to follow a circular
path, circumventing the MFR, to reach the location labeled by X. The targets
trajectory following the sequence of maneuvering models as shown in Fig. 11
is illustrated in this figure.

Fig. 11. The sample path of maneuvering models obtained from the discrete
stochastic approximation algorithm.

distance from the MFR stays fixed along the circular path, the
MFRs transition between modes depends only on the targets
maneuvering models. The SCFGs, because they correspond to
the micro control, are identical across the modes (the grammar
used here is the same as the one used previously), but the
transition matrix of the radar manager varies depending on the
targets maneuvering model. In this scenario, the simulation
results from both algorithms look virtually identical, and only
one set of results will be presented. Fig. 11 illustrates a sample
path of the maneuvering models obtained from the algorithm,
and Fig. 10 is the flight trajectory of the target following the maneuvering models. It can be seen that high maneuvering models

1118

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 56, NO. 3, MARCH 2008

SCFGs are multitype Galton-Watson branching processes, the

algorithms proposed in this paper can be viewed as filtering
and estimation of a partially observed multitype Galton-Watson
branching processes.
APPENDIX
A. A Justification of Logit Model
The Logit model can be justified by utility maximization argument. Consider only binary Logit model for simplicity, the
utilities of the decisions (advancing up or down the state space
as illustrated in Fig. 5) are

Fig. 12. Empirical distribution of the occupancies in the four maneuvering

models.

are deployed at the end to ensure its survivability. Fig. 12

shows the empirical distribution of the mode occupancies after
running the algorithms for 10 times, and it is observed that the
maneuvering model with the highest empirical distribution is
the one with the least threat, i.e., least average tracking time.
One implementation detail of the algorithm is the initialization of the Markov chain and the SCFGs. The initial parameter
values are fixed for each computation of the cost function because the stochastic approximation algorithm requires the estimator to be consistent. The Markov chain is initialized uniformly, and the SCFG is initialized according to the pretraining
method introduced in [28]. Briefly, the training data is first used
to train a hidden Markov model with start and terminating states.
The trained HMM is mapped to its approximated SCFG counterpart, and that is used as the initial configuration for the SCFG.

where is random threshold value. The threshold value indicates the amount of threat the MFR could take before switching
of states is desired. The threshold value is random because different targets may have different threshold values. Assuming
that the MFR always selects the decision with the highest utility,
the probability of going up in state can be expressed as

Suppose that the random variable has the logistic distribution,

the probability of advancing up the states, under the utility maximization argument, is expressed as

A more general discussion for more than two states can be found
in [49].

VI. CONCLUSION

ACKNOWLEDGMENT

The main idea of this paper is to model and characterize MFR

as a string generating device, where the control rules are specified in terms of SCFG modulated by the radars current tactical goal, and which is modeled by a Markov chain. This is unlike modeling of targets, where hidden Markov and state space
models are adequate [22], [23]. The modeling is knowledge
based, where each production rule corresponds to a operational
rule employed by the MFR to generate its radar words, and such
domain specific knowledge is assumed to be supplied by expert
radar analysts. The signal interpretation of the MFR, under our
formulation, is reduced to a state estimation by parsing through
radar words, and a maximum likelihood sequence estimator is
derived to evaluate the threat poses by the MFR. A maximum
likelihood parameter estimator is also derived to infer the unknown model parameters with the Expectation Maximization
algorithm. In addition, based on the interpreted radar signal,
the interaction dynamics of the MFR and the target is studied
and the control of the aircrafts maneuvering models is formulated as a discrete stochastic approximation problem. Since

The authors would like to thank Dr. F. Dilkes and Dr.

P. Lavoie of the Defense Research and Development Canada
for providing useful feedback on the material presented in
this paper. They would like to thank Dr. Dilkes for providing
the specification of the Mercury emitter that was used as an
example of modeling of a realistic MFR.
REFERENCES
[1] R. G. Wiley, Electronic Intelligence: The Analysis of Radar Signals.
Norwood, MA: Artech House, 1993.
[2] N. J. Whittall, Signal sorting in ESM systems, in Inst. Elect. Eng.
Proc. F, 1985, vol. 132, pp. 226228.
[3] D. R. Wilkinson and A. W. Watson, Use of metric techniques in
ESM data processing, in Inst. Elect. Eng. Proc. F, 1985, vol. 132, pp.
229232.
[4] J. Roe and A. Pudner, The real-time implementation of emitter identification for ESM, in Inst. Elect. Eng. Colloq. Signal Process. Electron.
Warefare, 1994, pp. 7/17/6.
[5] J. A. V. Rogers, ESM processor system for high pulse density
radar environments, in Inst. Elect. Eng. Proc. F, 1985, vol. 132, pp.
621625.
[6] M. I. Skolnik, Introduction to Radar Systems. New York: McGrawHill, 2002.

WANG AND KRISHNAMURTHY: SIGNAL INTERPRETATION OF MFRS

[7] S. Sabatini and M. Tarantino, Multifunction Array Radar System Design and Analysis. Norwood, MA: Artech House, 1994.
[8] N. Chomsky, Three models for the description of language, IRE
Trans. Inf. Theory, vol. 2, no. 3, pp. 113124, 1956.
[9] N. Chomsky and G. A. Miller, Finite state languages, Inf. Contr., vol.
1, no. 2, pp. 91112, May 1958.
[10] N. Chomsky, On certain formal properties of grammars, Inf. Contr.,
vol. 2, no. 2, pp. 137167, Jun. 1959.
[11] N. Chomsky, A note on phrase structure grammars, Inf. Contr., vol.
2, no. 4, pp. 393395, Dec. 1959.
[12] K. S. Fu, Syntactic Pattern Recognition and Applications. Englewood
Cliffs, NJ: Prentice-Hall, 1982.
[13] Y. A. Ivanov and A. F. Bobick, Recognition of visual activities and
interactions by stochastic parsing, IEEE Trans. Pattern Anal. Machine
Intell., vol. 22, pp. 852872, 2000.
[14] W. Zhu and J. Garcia-Frias, Modeling of bursty channels using stochastic context-free grammars, in Proc. 55th Veh. Technol. Conf. (VTC
2002), Birmingham, AL, May 2002, vol. 1, pp. 355359.
[15] W. Zhu and J. Garcia-Frias, Stochastic context-free grammars and
{Hidden Markov Models for modeling of bursty channels, IEEE
Trans. Veh. Technol., vol. 53, no. 3, pp. 666676, May 2004.
[16] P. Baldi and S. Brunak, Bioinformatics: The Machine Learning Approach, 2nd ed. Cambridge, MA: MIT Press, 2001.
[17] Y. Sakakibara, Grammatical inference in bioinformatics, IEEE
Trans. Pattern Anal. Machine Intell., vol. 27, pp. 10511062, 2005.
[18] R. Durbin, S. Eddy, A. Krogh, and G. Mitchison, Biological Sequence
Analysis: Probabilistic Models of Proteins and Nucleic Acids. Cambridge, U,K.: Cambridge Univ. Press, 1998.
[19] E. Rivas and S. R. Eddy, The language of RNA: A formal grammar
that includes pseudoknots, Bioinformatics, vol. 16, pp. 334340, 2000.
[20] M. Vilain, Getting serious about parsing plans: A grammatical analysis of plan recognition, in Proc. 8th Nat. Conf. Artif. Intell., 1990, pp.
190197.
[21] D. V. Pynadath and M. P. Wellman, Probabilistic state-dependent
grammars for plan recognition, in Proc. 16th Ann. Conf. Uncertainty
in Artif. Intell., 2000, pp. 507514.
[22] S. S. Blackman and R. Popoli, Design and Analysis of Modern Tracking
Systems. Norwood, MA: Artech House, 1999.
[23] Y. Bar-Shalom and X. R. Li, Estimation and Tracking: Principles,
Techniques, and Software. Norwood, MA: Artech House, 1993.
[24] N. A. Visnevski, F. A. Dilkes, S. Haykin, and V. Krishnamurthy, Nonself-embedding context-free grammars for multi-function radar modelingElectronic warfare application, in Int. Radar Conf., 2005, pp.
669674.
[25] N. A. Visnevski, V. Krishnamurthy, A. Wang, and S. Haykin, Syntactic modeling and signal processing of multifunction radars: A stochastic context free grammar approach, Proc. IEEE, 2007, submitted
for publication.
[26] A. V. Aho and J. D. Ullman, The Theory of Parsing, Translation
and Compiling. Englewood Cliffs, NJ: Prentice-Hall, 1972, vol. I,
Parsing.
[27] J. E. Hopcroft, R. Motwani, and J. D. Ullman, Introduction to
Automata Theory, Languages, and Computation. Reading, MA:
Addison-Wesley, 2001.
[28] K. Lari and S. J. Young, The estimation of stochastic context free
grammars using the Inside-Outside algorithm, Comp. Speech Language, vol. 4, pp. 3556, 1990.
[29] M. I. Miller and A. OSullivan, Entropies and combinatorics of
random branching processes and context-free languages, IEEE Trans.
Inf. Theory, vol. 38, pp. 12921310, 1992.
[30] C. G. Cassandras and S. Lafortune, Introduction to Discrete Event Systems. Boston, MA: Kluwer Academic, 1999.
[31] J. K. Baker, Trainable grammars for speech recognition, in Speech
Commun. Papers for the 97th Meet. Acoust. Soc. Amer., 1979, pp.
547550.
[32] D. C. Schleher, Electronic Warfare in the Information Age. Norwood,
MA: Artech House, 1999.
[33] P. L. Bogler, Radar Principles With Applications to Tracking Systems. New York: Wiley, 1990.
[34] R. Durbin, S. Eddy, A. Krogh, and G. Mitchison, Biological Sequence
Analysis: Probabilistic Models of Proteins and Nucleic Acids. Cambridge, U.K.: Cambridge Univ. Press, 1998.
[35] C. D. Manning and H. Schtze, Foundations of Statistical Natural Language Processing. Cambridge, MA: The MIT Press, 1999.
[36] A. Wang, V. Krishnamurthy, F. A. Dikes, and N. A. Visnevski, Threat
estimation by electronic surveillance of multifunction radars: A
stochastic context free grammar approach, in Conf. Decision Contr.,
2006.

1119

[37] X. R. Li and V. P. Jilkov, Survey of maneuvering target tracking. Part

v: Multiple-model methods, IEEE Trans. Aerosp. Electron. Syst., pp.
12551321, 2005.
[38] S. S. Blackman, M. T. Busch, and R. F. Popoli, IMM/MHT tracking
and data association for benchmark tracking problem, in Proc. Amer.
Contr. Conf., 1995, pp. 26062610.
[39] M. T. Busch and S. S. Blackman, Evaluation of IMM filtering for an
air defense system application, in Proc. SPIE, 1995, vol. 2561, pp.
435447.
[40] R. A. Singer, Estimating optimal tracking filter performance for
manned maneuvering targets, IEEE Trans. Aerosp. Electron. Syst.,
vol. AES-5, pp. 473483, 1970.
[41] J. L. Gertz, Multisensor surveillance for improved aircraft tracking,
The Lincoln Lab. J., vol. 2, pp. 381396, 1989.
[42] N. A. Visnevski, S. Haykin, V. Krishnamurthy, F. A. Dilkes, and P.
Lavoie, Hidden markov models for radar pulse train analysis in electronic warfare, in IEEE Int. Conf. Acoust., Speech, Signal Process.,
2005, pp. 597600.
[43] Z. Chi, Statistical properties of probabilistic context-free grammars,
Computat. Linguist., vol. 25, pp. 131160, 1999.
[44] L. R. Rabiner, A tutorial on hidden Markov models and selected applications in speech recognition, Proc. IEEE, vol. 77, pp. 257286,
1989.
[45] A. P. Dempster, N. M. Laird, and D. B. Rubin, Maximum likelihood
from incomplete data via the EM algorithm, J. Roy. Statist. Soc., vol.
39, pp. 138, 1977.
[46] C. F. J. Wu, On the convergence properties of the algorithm, Ann.
Statist., vol. 11, no. 1, pp. 95103, 1983.
[47] D. Goldsman and B. L. Nelson, Statistical screening, selection, and
multiple comparison procedures in computer simulation, in Proc.
1998 Winter Simulation Conf., 1998.
[48] V. Krishnamurthy, X. Wang, and G. Yin, Spreading code optimization
and adaptation in CDMA via discrete stochastic approximation, IEEE
Trans. Inf. Theory, pp. 19271949, 2004.
[49] J. S. Cramer, Logit Models From Economics and Other Fields. Cambridge, U.K.: Cambridge Univ. Press, 2003.

Alex Wang was born in 1979 in Taiwan. He received

the B.S. degree (with honors) in engineering physics
with a commerce minor and the M.S. degree in electrical and computer engineering from the University
of British Columbia in 2003 and 2005, respectively.
He is currently pursuing the Ph.D. degree in
statistical signal processing, under the supervision
of Dr. V. Krishnamurthy, at the University of British
Columbia, Vancouver, Canada. His research interests
include radar signal processing, syntactic pattern
recognition, and uncertain reasoning.

Vikram Krishnamurthy (S90-M91-SM99-F05)

was born in 1966. He received the B.S. degree from
the University of Auckland, New Zealand, in 1988
and the Ph.D. degree from the Australian National
University, Canberra, in 1992.
Since 2002, he has been a professor and Canada
Research Chair of the Department of Electrical Engineering, University of British Columbia, Vancouver,
Canada. Prior to 2002, he was a chaired professor
with the Department of Electrical and Electronic Engineering, University of Melbourne, Australia, where
he also served as Deputy Head of department. His current research interests include stochastic modeling of biological ion channels, stochastic optimization
and scheduling, and statistical signal processing.
Dr. Krishnamurthy has served as Associate Editor for several journals including the IEEE TRANSACTIONS ON SIGNAL PROCESSING, IEEE
TRANSACTIONS AEROSPACE AND ELECTRONIC SYSTEMS, IEEE TRANSACTIONS
CIRCUITS AND SYSTEMS B, and IEEE TRANSACTIONS NANOBIOSCIENCE, AND
SYSTEMS AND CONTROL LETTERS. He is coeditor with S. H. Chung and O.
Andersen of Biological Membrane Ion ChannelsDynamics Structure and
Applications (New York: Springer-Verlag, 2006).